For the complete documentation index, see llms.txt. This page is also available as Markdown.

AI prompts

Control how the AI writes — brand voice, rules, and which of your product data it sees.

Every AI generator in Product Pelican runs on a prompt you can edit. The AI Prompts page is where you set your brand voice, add rules the AI must follow, and choose which parts of your product data it gets to see.

📸 Screenshot needed — the AI Prompts page with the prompt list on the left and an editor open on the right.

The prompts

Prompts are grouped by what they apply to.

Products

Prompt
What it generates
Where it's used

Title

An improved, search-friendly product title

Products → Titles

Description

The product description

Products → Descriptions

Metafields

Values for metafields you choose, generated alongside the description

Products → Descriptions

Image Alt Text

A description of a product image

Products → Missing Alt Text

Category

The best-matching Shopify taxonomy category — and, optionally, the matching Google product category

Products → No Category

Category Attributes

Category-specific attributes — Colour, Size, Material…

Products → Missing Category Attributes

SEO

The SEO title and meta description

Products → SEO

FAQ Generation

Question-and-answer pairs

Product FAQ

Collections

Prompt
What it generates
Where it's used

Description

A collection description

Collections → Description

Banner Alt Text

A description of a collection banner image

Collections → Missing Banner Alt Text

SEO

The SEO title and meta description for a collection

Collections → SEO

FAQ Generation

Question-and-answer pairs for a collection page

Product FAQ → Collections

How prompts work

A prompt is plain instructions, not a template. You do not need placeholders or variable syntax — Product Pelican attaches the product data automatically before sending your instructions to the AI.

Each prompt's editor lists exactly what gets attached. For example, the description prompt automatically receives the product title, type, tags, the current description, any context metafields you've ticked, and — when image analysis is on — the product photos.

So write your prompt as direction:

Write in a warm, plain-spoken voice. Lead with what the product is and who it's for. Mention materials and care in the second paragraph. British English. No exclamation marks, no "elevate", no "curated". Two short paragraphs, under 120 words.

Rather than trying to describe the data.

What you can and can't change

Prompts replace the instruction and persona part of the request. The structural parts — the candidate category list, the required output format — are always added by the app, so a customised prompt cannot break the pipeline. Rewriting the category prompt changes how the AI chooses; it cannot make it return something the app can't read.

Feeding your own data in

Several prompts have a context metafield picker — a checklist of your product metafield definitions. Anything you tick is passed to the AI as additional context.

Available on: Title, Description, Metafields, and FAQ Generation.

This is how you get genuinely specific output. If you keep fabric composition, care instructions, dimensions or warranty terms in metafields, tick them, and generated copy will reference real facts instead of generalising from the title.

📸 Screenshot needed — the context metafield picker with several product metafields ticked.

Generating metafields alongside descriptions

The Metafields prompt is different from the others: instead of writing one field, it fills a set of metafields you nominate, in the same pass as the description.

For each metafield you configure, you give it:

  • the metafield to write to

  • its own prompt describing what should go in it

Then, whenever you generate descriptions, those metafields are generated too. In the results, the description and each metafield get their own accept checkbox, so you can take the metafields and reject the description or vice versa. Any configured field the AI returns empty is listed explicitly rather than silently skipped.

Google Shopping fields

The same pane has a Google Merchant Center section: a ready-made set of the fields a Google feed needs — condition, gender, age group, custom product and five custom labels — that you tick rather than having to create metafields for.

They generate and apply exactly like the metafields above, with one difference: fields with a fixed set of values (condition, gender, age group) are shown as a dropdown, and an answer outside Google's vocabulary is dropped rather than written.

The Google product category is separate again — it is switched on from the Category prompt and is a lookup rather than a generation. Full detail: Google Shopping fields.

📸 Screenshot needed — the Google Merchant Center section with condition and age group ticked.

This is the fastest way to populate a PIM. One pass over your catalogue fills the description and every structured attribute you care about, from the same source data and the same photos.

Reviewing what the AI was thinking

Generated descriptions and metafields carry a Show thinking link exposing the reasoning behind the result.

It is worth reading when:

  • A result surprises you and you want to know which input caused it.

  • A generation failed — the transcript is still returned, and usually explains why.

  • You are tuning a prompt and want to see whether an instruction is landing.

Tips

  • Change one thing at a time. Generate for the same 5 products after each edit so you can attribute the difference.

  • Be specific about length. "Two short paragraphs, under 120 words" works; "concise" does not.

  • Ban your least favourite words explicitly. A list of forbidden phrases is more effective than describing a tone.

  • Test on your weirdest products. The ones with sparse data are where prompts break down.

  • Turn on image analysis before tuning description prompts — it changes the output substantially, so tuning without it means tuning twice.

Last updated